Analysing longitudinal wearable physical activity data using non-stationary time series models.

Area of Science:

• Biomedical Engineering
• Data Science
• Statistics

Background:

• Wearable devices offer new ways to monitor physical activity.
• Traditional methods often ignore temporal changes, focusing on averages or snapshots.
• A novel statistical method is needed to analyze longitudinal physical activity data, accounting for temporal structure.

Purpose of the Study:

• To develop and validate a novel statistical method for analyzing longitudinal physical activity data from wearable devices.
• To account for the temporal structure and dynamics within the data.
• To provide enhanced understanding of physical activity patterns over time.

Main Methods:

• Utilized secondary data from the Multidimensional Individualised Physical Activity (MIPACT) randomized controlled trial.
• Applied a Trend Locally Stationary Wavelet (TLSW) model to analyze 12-week physical activity data from 80 participants.
• Introduced the Time in Reference Region of Variability (TIRRV) metric to assess individual changes relative to baseline.

Main Results:

• Demonstrated the effectiveness of the TLSW approach for analyzing hourly resolution physical activity data.
• The TLSW model accounts for time dependency and data structure, enabling detailed trend and confidence interval analysis.
• The TIRRV metric provides a baseline-informed assessment of individual and group progress over time, yielding robust insights.

Conclusions:

• The TLSW-based approach is a novel and effective method for analyzing high-resolution physical activity data from wearable technology.
• TLSW trends robustly characterize individual and group behavior over extended periods.
• This approach enhances the understanding of temporal changes in device-measured physical activity for researchers, clinicians, and patients.